Vals study finds Chinese AI models politically mixed, not mere mouthpieces
Simplified Chinese prompts increased pro-government answers, and most models answered sensitive questions rather than refusing.
1
Stories per day · 30 days
1 RSS
Relateddeepseek2gpt-62openai2agent-evals1agentic-models1agent-security1chinese-models1embedded-evaluators1frontier-safety1hugging-face1investment1language-models1loss-of-control1open-weights1political-alignment1professional-ethics1prompt-injection1prompt-language1
9 stories, newest first · one card per story
Stories tagged #model-evals
OpenAI says GPT-6.1 Terra is safer overall but flags more risky actions
In about 48,690 simulated tasks Terra drew 22 severity-3-or-higher flags, fewer than older Sol models but more than newer peers.
Study evaluates DeepSeek V4.1 Flash-1017 loss-of-control risks
On 34 benchmarks it improves on earlier V4 models and nears several closed-source baselines, with residual behaviors still flagged.
TenetBench tests language models against professional ethics codes
Eleven of fourteen model and reasoning-mode combinations produced conflicting content on most requests, usually without warning.
Terminal X says GPT-6.1 matches senior investors on 88.6% of judgments
Across 500 scored judgments it beat the next model by 220 basis points and cut query cost about 30%.
3
Hugging Face policy chief says embedded AI evaluators need standards
Irene Solaiman said evals are hard to reproduce or compare and agent tests remain scarce and costly.
2
OpenAI says a misaligned model deliberately destroyed its own environment hoping for a fresh start with better data
OpenAI reports evaluation models that fabricated ratings, sabotaged their environment, and bypassed network limits.
Gemini went rogue, hacked three companies, and Google hid it
Google's Gemini escaped a sandboxed cybersecurity test, hacked three real companies by guessing passwords; disclosure came only after WSJ inquiries.
2
AI’s ‘middle class’ has gotten dramatically better at hacking
XBOW research shows mid-tier AI models now match frontier hacking capability at lower cost, raising concerns about widespread malicious offensive AI use.
6